pig-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Apache Wiki <wikidi...@apache.org>
Subject [Pig Wiki] Update of "PigExecutionModel" by ChrisOlston
Date Fri, 04 Jan 2008 19:25:36 GMT
Dear Wiki user,

You have subscribed to a wiki page or wiki category on "Pig Wiki" for change notification.

The following page has been changed by ChrisOlston:

New page:
---+ Pig Execution Model

---++ Introduction

The goal is to decide how to structure operator evaluation pipelines in Pig. The major issues
include whether data flow follows a push model or a pull model, whether an operator evaluation
tree is multi-threaded or single-threaded, and what the API for user-defined functions (UDFs)
looks like.

---++ Alternative Execution Models

Some possible execution models are:

---+++ Model 1

   * One thread.
   * Top-level scheduler round-robins through "leaf" operators.
   * Each time an operator is invoked, it gets handed exactly one tuple.
   * Special "split" operator buffers data that gets fed to multiple operators; at most one
tuple gets buffered at each split point.
   * UDF API: declare zero or one input bag as "streaming"; init() hands it all data except
the streaming bag; next() hands one tuple from streaming bag

---+++ Model 2

   * One thread per "leaf" operator.
   * Scheduling done by OS.
   * Operator gets to read as many tuples as it wants; if it reads from multiple inputs, can
interleave "next()" calls on the inputs in arbitrary fashion.
   * Split operator may buffer up to K tuples (or B bytes); if an operator tries to read too
far ahead it gets blocked until other operators reading from the same buffer catch up.
   * Deadlock can arise; need to detect it and release it by relaxing the K/B constraint on
one or more of the split buffers.

---+++ Discussion

---++++ Model 1 Drawbacks

   * underutilize multi-core systems? (depends whether the policy is to assign several map/reduce
tasks to a machine)
   * difficult (or impossible?) to support operations that require streaming access to multiple
inputs (e.g., merge join, merge-based set difference, etc. which operate over pre-sorted input
   * UDF APIs more complex?

---++++ Model 2 Drawbacks

   * thread synchronization overhead
   * complexity of multi-threaded implementation

---++++ Related work

   * Fjords paper [[http://db.lcs.mit.edu/madden/html/madden_fjords.pdf][paper]][[http://www.cs.umd.edu/class/fall2002/cmsc818s/Lectures/fjords.pdf][slides]]
   * Stream Programming Model / MIT StreamIt [waiting for pointer]

View raw message